(SDXL) Fix --ondemand and vae scale factor use, and fix VAE flags.

apps/stable_diffusion/src/models/model_wrappers.py

@@ -165,13 +165,8 @@ def __init__(
         self.check_params(max_len, width, height)
         self.max_len = max_len
         self.is_sdxl = is_sdxl
-        self.height = height
-        self.width = width
-        if is_sdxl:
-            # We need to scale down the height/width by vae_scale_factor, which
-            # happens to be 8 in this case.
-            self.height = height // 8
-            self.width = width // 8
+        self.height = height // 8
+        self.width = width // 8
         self.batch_size = batch_size
         self.custom_weights = custom_weights.strip()
         self.use_quantize = use_quantize

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_utils.py

@@ -33,6 +33,7 @@
     end_profiling,
 )
 import sys
+import gc
 from typing import List, Optional
 
 SD_STATE_IDLE = "idle"
@@ -189,6 +190,7 @@ def load_vae(self):
     def unload_vae(self):
         del self.vae
         self.vae = None
+        gc.collect()
 
     def encode_prompt_sdxl(
         self,
@@ -327,6 +329,7 @@ def encode_prompt_sdxl(
 
         if self.ondemand:
             self.unload_clip_sdxl()
+            gc.collect()
 
         # TODO: Look into dtype for text_encoder_2!
         prompt_embeds = prompt_embeds.to(dtype=torch.float32)
@@ -387,6 +390,7 @@ def encode_prompts(self, prompts, neg_prompts, max_length):
         clip_inf_time = (time.time() - clip_inf_start) * 1000
         if self.ondemand:
             self.unload_clip()
+            gc.collect()
         self.log += f"\nClip Inference time (ms) = {clip_inf_time:.3f}"
 
         return text_embeddings
@@ -499,6 +503,8 @@ def produce_img_latents(
         if self.ondemand:
             self.unload_unet()
             self.unload_unet_512()
+            gc.collect()
+
         avg_step_time = step_time_sum / len(total_timesteps)
         self.log += f"\nAverage step time: {avg_step_time}ms/it"
 
@@ -556,6 +562,8 @@ def produce_img_latents_sdxl(
                 break
         if self.ondemand:
             self.unload_unet()
+            gc.collect()
+
         avg_step_time = step_time_sum / len(total_timesteps)
         self.log += f"\nAverage step time: {avg_step_time}ms/it"
 
@@ -652,7 +660,6 @@ def from_pretrained(
                 use_lora,
                 ondemand,
             )
-
         return cls(scheduler, sd_model, import_mlir, use_lora, ondemand)
 
     # #####################################################
@@ -765,6 +772,7 @@ def encode_prompts_weight(
         clip_inf_time = (time.time() - clip_inf_start) * 1000
         if self.ondemand:
             self.unload_clip()
+            gc.collect()
         self.log += f"\nClip Inference time (ms) = {clip_inf_time:.3f}"
 
         return text_embeddings.numpy()

apps/stable_diffusion/src/utils/utils.py

@@ -565,9 +565,10 @@ def get_opt_flags(model, precision="fp16"):
         iree_flags += opt_flags[model][is_tuned][precision][
             "specified_compilation_flags"
         ][device]
-    # Due to lack of support for multi-reduce, we always collapse reduction
-    # dims before dispatch formation right now.
-    iree_flags += ["--iree-flow-collapse-reduction-dims"]
+    if "vae" not in model:
+        # Due to lack of support for multi-reduce, we always collapse reduction
+        # dims before dispatch formation right now.
+        iree_flags += ["--iree-flow-collapse-reduction-dims"]
     return iree_flags
 
 

apps/stable_diffusion/web/ui/txt2img_sdxl_ui.py

@@ -147,6 +147,8 @@ def txt2img_sdxl_inf(
         # For SDXL we set max_length as 77.
         print("Setting max_length = 77")
         max_length = 77
+        if global_obj.get_cfg_obj().ondemand:
+            print("Running txt2img in memory efficient mode.")
         txt2img_sdxl_obj = Text2ImageSDXLPipeline.from_pretrained(
             scheduler=scheduler_obj,
             import_mlir=args.import_mlir,
@@ -164,7 +166,7 @@ def txt2img_sdxl_inf(
             debug=args.import_debug if args.import_mlir else False,
             use_lora=args.use_lora,
             use_quantize=args.use_quantize,
-            ondemand=args.ondemand,
+            ondemand=global_obj.get_cfg_obj().ondemand,
         )
         global_obj.set_sd_obj(txt2img_sdxl_obj)